Geographic Targeting Models

Spatial data is a goldmine, but how you engineer and select those features can make or break your machine learning model. Here is my F to S Tier list of using spatial features in Data Science: F-Tier: Raw Lat/Lon into Tree-Based Models - Feeding raw Latitude and Longitude directly into an XGBoost or Random Forest model. Trees make orthogonal splits, meaning they struggle to isolate specific locations. Not much comes of this. D-Tier: One-Hot Encoding ZIP Codes - Turning 10,000 ZIP codes or neighborhoods into dummy variables. Zip code numbers tell you nothing about the areas they represent (and are postal routes not actual boundaries) and you have completely erased the actual physical distance between those locations. C-Tier: Distance to Points of Interest (POIs) Calculating the Haversine or Manhattan distance to key landmarks (city centers, transit stops, competitors). This is a better, a reliable baseline that actually gives your model actionable context about where something is relative to what matters. B-Tier: Spatial Indexing (Geohash / H3) - Converting coordinates into localized bins like A5 or Uber’s H3 Hexagons. This allows you to aggregate historical target variables by region and captures non-linear spatial relationships without blowing up the feature space. A-Tier: Spatial Lags & Graph Embeddings - Encoding the "neighborhood effect." Instead of just looking at the point itself, you calculate Spatial Lags (e.g., the average target value of the 5 nearest neighbors) or use Graph Neural Networks (GNNs) to create spatial embeddings. This explicitly models the rule that things close together are more related than things far apart. AKA use PySAL. S-Tier: Spatial Regression for Feature Selection - Standard feature importance metrics (like SHAP or tree-based Gini) will lie to you if spatial autocorrelation is present. Using spatial regression technique like Geographically Weighted Regression (GWR) or Spatial Error Models (SEM) to filter your features ensures you aren't just keeping redundant variables that act as secret proxies for location. It is the ultimate reality check for your spatial dataset. But let me know what you do with your data - would love to hear other ideas. 🌎 I'm Matt Forrest and I talk about modern GIS, earth observation, AI, and how geospatial is changing. 📬 Want more like this? Join 12k+ others learning from my daily newsletter → forrest.nyc

There’s more to geo testing than matched markets A common misconception is that matched market tests are the only way to do geo testing. Economists and statisticians have made enormous progress in geo testing so you don’t have to settle and can do regional experiments instead! Matched market tests, for example, are when you do something in Detroit and compare it to Milwaukee. Regional experiments, on the other hand, randomly sample a larger number of markets into control and treatment groups. Matched market methods have limitations compared to regional experiments: 1. Less precision. With only 1-3 locations in the treated zone, the error on any matched market test analysis is going to be much higher than it would be for a broader test that can average out noise across many regions. 2. Less transferable insights. Different regions may respond differently to the same creatives and campaign configurations. As you isolate your treatment down to 1-3 locations for a matched market test you lose credibility that the estimated impacts for those areas will transfer to the rest of the country. 3. Downward bias on estimated ROI. Depending on how much you plan to spike your marketing for the matched markets, you may be flooding them with much more spend than you otherwise would have in a business as usual setting. If that's the case, you will be further down the diminishing marginal return curve than in business as usual, making your overall ROI look smaller. The most accurate and precise regional experiments use frontier methods: * At Haus, instead of taking the naive average of a few matched markets, we build synthetic control models that weigh control regions to best fit the targeted regions. For example, for your business, Seattle might be comparable to 60% San Francisco, 30% Denver, and 10% San Diego. * And we don’t stop at traditional synthetic control, Haus has PhD scientists focused on continually improving the performance of the models we use. This is why science matters.

New #GeoAI paper out led by our fantastic students Haoran Zhang & Livia Betti and w/ Esther Rolf and David Alvarez Melis 🌐 🤖 🌎 Deploying a model trained on data from West Africa in East Asia can be difficult due to geographic distribution shifts - and the geographic distance between the two areas is somewhat indicative of that. We investigate whether embeddings of geospatial data can be used to strengthen this signal; and it turns out, they can! And not just image or text embeddings, we show that location embeddings obtained from pretrained location encoders (SatCLIP, GeoCLIP) are indicative of how well a model trained on area A will perform in area B - irrespective of the downstream task! As embeddings from location encoders are available for any location worldwide, this allows us to create a new, globally available distance metric quantifying geographic distribution shifts: GeoSpOT. GeoSpOT can be used to inform decisions such as (1) where to sample training data for a model that targets a specific area or (2) to which area a model trained on a specific region will transfer well to. Read the full paper and check our code 👇 📰 Paper: https://lnkd.in/gFigHfsC 💻 Code: https://lnkd.in/gGxAeDa6

🎯 The Power of Geo-Targeting in Search Arbitrage Success 🌎💰 Search arbitrage is all about driving paid traffic at a lower cost than the revenue it generates. But here’s a game-changer: geo-targeting. Not all locations perform the same. A $0.10 CPC in one state could be $1.50 in another, and conversion rates can be even more unpredictable. If you’re running arbitrage campaigns without a geo-specific approach, you’re leaving money on the table. 📍 Why the U.S. Market is Different? The U.S. is a high-value but competitive market. Advertisers bid aggressively, and user behavior varies state by state. A campaign that thrives in California might burn cash in Texas. 🔹 USA = High Ad Costs, High Monetization Potential 🔹 Other Countries = Lower Ad Costs, Lower Monetization But even within the U.S., not all states are equal. This is where state-by-state targeting comes in. ⚠️ How Often Should You Change Geo-Targeting in the U.S.? On Facebook (Meta Ads), the worst mistake is running the same geo-targeting for too long without optimizing. 📊 The rule of thumb? Rotate and test every 2-5-10 days. 🚀 Pro Tip: If CTR or conversions drop, pause low-performing states and test new ones. 🌎 Best Niches & State-Specific Insights Different states have different user behaviors, incomes, and ad costs. Some niches explode in specific regions while failing elsewhere. ✅ California – Tech, Solar Energy, Luxury Real Estate ➡ High-income users, but expensive clicks. Great for high-ticket offers like EVs, SaaS, or premium real estate leads. ✅ Washington – Eco-Friendly Products, Software, Finance ➡ Strong tech presence (Seattle), eco-conscious audience. Good for fintech, green energy, and investment niches. ✅ Texas – Home Improvement, DIY, Insurance ➡ Huge demand for roofing, HVAC, and auto insurance. Lower ad costs than CA, but massive scale potential. ✅ Florida – Retirement, Healthcare, Travel ➡ Older population. Great for Medicare ads, real estate, and travel offers. ✅ New York – Luxury, Finance, High-Ticket Services ➡ Expensive but profitable for B2B, legal, and wealth management niches. 🔥 Final Takeaway: Let the Data Guide You Don’t stick to one state blindly. Let CPC, conversion rate, and ad revenue dictate your next move. 💡 Actionable Steps: ✅ Start broad (test multiple states). ✅ Pause underperforming regions after 5-10 days. ✅ Scale in states with low CPC & high conversions. ✅ Always A/B test state-specific ad creatives. Geo-targeting isn’t just a setting, it’s a scaling strategy. If you’re still treating all U.S. traffic the same, it’s time to rethink your approach. 🚀 Have you experimented with geo-targeting in arbitrage? Drop your insights below! 👇 #SearchArbitrage #DigitalMarketing #FacebookAds #GeoTargeting #GrowthHacking #AffiliateMarketing #AdOptimization

#Stop Mapping Rocks. #Start Targeting Mineral Systems. Most exploration programs fail at the same point: they map lithologies, not ore-forming systems. Modern mineral targeting is not descriptive, it is predictive. It focuses on identifying where structure, fluid flow, and chemical reactivity converge to concentrate economic mineralization. I use a mineral systems approach that consistently improves target quality by focusing on> • #Structural Controls That Matter Fault intersections, shear zones, fold hinges, and lithological contacts that act as fluid conduits and ore traps not just mapped features. • #Hydrothermal Alteration Vectors Silicification, sericitization, chloritization, and iron oxide zones used as direct footprints toward mineral centers, not background geology. • #Active Plumbing Systems Quartz-carbonate-sulfide vein networks, stockworks, and breccias that define active or fossilized hydrothermal flow pathways. • #Kinematic Reconstruction S–C fabrics, σ/δ porphyroclasts, mica fish, and Riedel shear systems used to decode fluid flow direction and deformation kinematics. • #Vector-Based Targeting Strategy Systematic progression from weak alteration → structural intensification → vein density increase → sulfide concentration = defined exploration vector toward ore zones. • #Multi-Scale Integration From regional tectonic architecture to outcrop-scale textures because ore bodies are not random; they are the product of linked crustal-scale processes. The outcome is simple> {Better geological understanding = fewer drill holes wasted = higher probability of discovery} [I work with exploration teams to refine targeting models, reduce geological uncertainty, and convert complex structural systems into actionable drill targets across gold, base metals, and critical mineral projects. Open to technical collaboration, consulting, and project-based engagement]

Location Targeting: The "Spray and Pray" Mistake "We target the entire metropolitan area because we serve all neighborhoods." This common approach is the equivalent of using a fire hose to water a flower garden. For a local service business, we analyzed performance data by location and found: 📌 Certain zip codes converted 6X better than others 📌 Some neighborhoods had 70% lower cost per lead 📌 Competition density varied dramatically by location Yet they were allocating budget equally across all areas. Our precision targeting approach: ➡️ Created separate campaigns for high-performing regions ➡️ Applied custom bid adjustments based on conversion data ➡️ Developed neighborhood-specific ad copy and extensions ➡️ Built location-based landing pages mentioning specific areas The result? A 52% increase in qualified leads without increasing budget. Smart geographic targeting isn't about maximum reach—it's about strategic investment where returns are highest. #LocalSEO #GeographicTargeting #GoogleAdsStrategy #MarketingROI

𝐔𝐬𝐢𝐧𝐠 𝐃𝐢𝐬𝐭𝐚𝐧𝐜𝐞 𝐌𝐞𝐚𝐬𝐮𝐫𝐞𝐬 𝐭𝐨 𝐏𝐫𝐞𝐝𝐢𝐜𝐭 𝐌𝐨𝐝𝐞𝐥 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐀𝐜𝐫𝐨𝐬𝐬 𝐆𝐞𝐨𝐠𝐫𝐚𝐩𝐡𝐢𝐜 𝐑𝐞𝐠𝐢𝐨𝐧𝐬: 𝐀 𝐏𝐫𝐚𝐜𝐭𝐢𝐜𝐚𝐥 𝐀𝐩𝐩𝐫𝐨𝐚𝐜𝐡 𝐟𝐨𝐫 𝐄𝐚𝐫𝐭𝐡 𝐎𝐛𝐬𝐞𝐫𝐯𝐚𝐭𝐢𝐨𝐧 When developing machine learning models for satellite imagery, we face a common dilemma: data is plentiful in some regions but scarce in others. How can we know in advance if a model trained on data-rich areas will work effectively when applied to regions with limited training data? Haoran Zhang et al. demonstrate that certain domain distance measures can serve as reliable predictors of how well models will transfer between geographic regions - even without labeled data from the target region. They found that when calculating distances between domain-specific data distributions (using image and location embeddings), larger distances typically correlate with greater performance drops during domain adaptation. This relationship held across different datasets, geographic domain definitions, and model architectures. This approach could improve how we deploy machine learning models across different geographic regions, helping to address the uneven distribution of data availability while ensuring models perform reliably where they're needed most. https://lnkd.in/dAd-JpYP #RemoteSensing #MachineLearning #ClimateChange #SpatialData #EarthObservation

Your ads are bringing in thousands of clicks from cities that will never buy from you. It’s called junk traffic. That's exactly what happened to a premium jewelry brand. Their campaigns looked successful on paper. 15,000 clicks monthly. 2.3% CTR. Low CPCs. But sales were almost break-even. We dug into their data: 67% of their traffic was coming from tier-3 cities and rural areas where purchasing power for ₹25K+ jewelry is extremely low. Users were clicking because the ads looked appealing, but conversion was impossible. Their targeting was set to "All of India" because they thought "bigger reach = more sales." Classic mistake. Here's how we fixed their targeting nightmare: 1/. Geographic Precision Surgery We analyzed their Shopify analytics and found 89% of actual sales came from just 12 cities: Mumbai, Delhi, Bangalore, Hyderabad, Chennai, Pune, Kolkata, Ahmedabad, Gurgaon, Noida, Chandigarh, and Jaipur. Excluded 200+ smaller cities that generated clicks but zero sales. Traffic dropped 60%, but conversion rate jumped 340%. 2/. Income-Based City Targeting Their products ranged from ₹15K to ₹1.5L. But they were targeting everyone in metro cities. We layered precise targeting: South Mumbai, South Delhi, Koramangala in Bangalore Excluded areas with lower purchasing power Focused on premium localities within metros 3/. Device and Platform Filtering Data showed mobile users from tier-2 cities had 0.1% conversion rates. Desktop users from metro areas had 4.8% conversion. We reallocated 70% of the budget to desktop targeting in high-value pin codes. 4/. Demographic and Interest Overlays English-speaking audiences from metros converted 5x better than regional language users for luxury jewelry. Added interest layers: luxury shoppers, wedding planning, premium brands, imported jewelry buyers. Results after 21 days: - Click costs increased 40% (but from quality audiences) Conversion rate jumped from 0.8% to 4.2% - ROAS improved from 1.1x to 5.7x Monthly revenue increased ₹3.4L with same ad spend The expensive clicks were actually cheap because they converted. Cliche, but helpful advice - Your audience quality matters more than audience quantity :) . . . . . . P.S. If you’re looking to scale your brand sustainably…book a call today for a free session. . . . . #performancemarketing #metaads

🌐📍 Here is how I transformed my Geo-Targeting Strategy with Google Sheets and Google Maps API 📍🌐 And I am all about providing value to the community. Imagine being able to effortlessly calculate the distance and driving time between multiple locations at once, all within a simple Google Sheet. ** No dev work required** With my Google Sheet's custom formulas powered by the App Script, connected to the Google Maps API, you can do just that! Whether you're looking to refine your geo-targeting strategy or manage bulk distance calculations, this has you covered! Here's how this tool can revolutionize your search engine management and digital marketing strategies. My favourite use case is the last one: ✅ Optimized Ad Targeting: Use distance calculations to get a better feel for how close your audience is and make super-targeted ad campaigns. This way, you can reach potential customers who are just around the corner, boosting engagement and CVRs. ✅ Cost-Effective Campaigns: Spot groups of high-value customers based on how close they are, and spend your ad budget more wisely. This helps cut down on wasted spend and boosts your return on ad spend (ROAS). ✅ Enhanced Local SEO: Tailor your local SEO strategies based on the geographic distribution of your customers. Optimize your Google My Business listings and local keywords for better visibility and rankings. ✅ Improved Customer Experience: Use driving duration data to set realistic delivery times and pickup windows. This makes customers happy, leading to better reviews and higher local search rankings. ✅ Data-Driven Insights: Check out patterns and trends in customer behavior to shape your bigger marketing strategies. ✅ Competitive Analysis: Compare driving times to your locations versus your competitors' to spot your logistical advantages. Highlight convenience in your marketing to attract more customers. ✅ Personalized Marketing: Use location data to create special offers. Offer in-store promotions for nearby customers and free shipping for those farther away, tailoring your approach to attract more interest. ✅ Event Planning: Plan events and in-person promotions at locations that maximize attendance and engagement, using distance and driving duration data to choose optimal spots. ✅ Understanding Customer Behavior for Physical Stores: For physical stores, understanding how far customers are willing to travel and their behavior is crucial. This data helps in tailoring marketing efforts, improving store accessibility, and planning inventory to meet local demand effectively. Are you ready to take your geo-targeting and digital marketing strategies to the next level? Comment "GEO" and I'll send over the Google Sheet and detailed instructions to get you started. ----- #googleads #sem #ppc #ppcstrategy #paidsearch 🔔 follow me for more paid search content ♻ spread the word